Guangdong Juyao Intelligent Technology Co., Ltd. · Laser Software & Process R&D
23°07′N 113°15′E · GUANGZHOU · PURE STATIC MONOGRAPH
← Back to research
Material & Process2026-11-02· 9 min read

Teaching a neural network to predict femtosecond groove geometry from 100 experiments: test R ≈ 0.95 — but the HAZ is only an optical proxy

Two inputs, two outputs, two hidden layers (10 + 5 neurons): predicting groove width and discoloration-band width together across 1–10 W and 10–100 mm/s, with 11–15% errors — valid as interpolation inside that parameter space only.

Microscope image of a single femtosecond-laser groove in aluminium foil: the dark groove flanked by a discoloration band, annotated with groove width and HAZ proxy width, 100 μm scale bar; machined at 10 W and 80 mm/s, measured groove width 71 μm and HAZ proxy width 120 μm (Fig. 2 of Tański et al., Materials 2026, CC BY 4.0; resized only)
COVERMicroscope image of a single femtosecond-laser groove in aluminium foil: the dark groove flanked by a discoloration band, annotated with groove width and HAZ proxy width, 100 μm scale bar; machined at 10 W and 80 mm/s, measured groove width 71 μm and HAZ proxy width 120 μm (Fig. 2 of Tański et al., Materials 2026, CC BY 4.0; resized only)
CORE RESEARCH QUESTION

Core question: with only laser power and scan speed as inputs, can a small neural network predict both the groove width and the thermal footprint of femtosecond grooving at once — and beat simple regression while doing it?

ABSTRACT

Tański and colleagues (Institute of Fluid Flow Machinery, Polish Academy of Sciences; Technical University–Sofia; Gdynia Maritime University) cut grooves into 200 μm thick aluminium 3003 foil with a femtosecond laser: 1030 nm, 500 fs, 200 kHz, a spot of about 100 μm, up to 10 W average power (50 μJ per pulse, up to 0.6 J/cm² on the sample), combining powers of 1–10 W (1 W steps) with scan speeds of 10–100 mm/s (10 mm/s steps) into 100 experiments. Optical microscopy measured groove width at ten positions per groove, and the total width of the visible discoloration band flanking each groove served as an optical proxy for the heat-affected zone (HAZ). A 2-10-5-2 feed-forward neural network (70% training, 15% validation, 15% testing, Levenberg–Marquardt training) predicted both outputs at once: test-set overall correlation R ≈ 0.95 (0.88 for groove width, 0.96 for the HAZ proxy width), with mean absolute percentage errors of about 11% for groove width and 15% for the HAZ proxy width — better than linear regression (0.74/0.66) and second-order polynomial regression (0.88/0.81) on the same data. This note checks the experimental grid, the network structure, the error structure and the baseline models, reconstructs the nonlinear “higher power and slower scan → wider groove, larger thermal footprint” relationship, and explains why the model cannot be lifted to other materials, other laser systems, or outside its parameter range. Cover: Fig. 2 of the paper (CC BY 4.0, resized only).

01 / Why groove prediction is harder than hole prediction

Femtosecond processing is often called “cold ablation”, but that label needs qualification: although each pulse is short, energy deposition involves nonlinear absorption, heat transfer, phase transitions, plasma formation and material ejection happening at once, so analytical models struggle to capture everything, and point-by-point trial-and-error experiments are slow and expensive. Machine learning is attractive here precisely because it skips solving all the physics and instead builds a usable mapping from parameters to outcomes.

Most earlier machine-learning work on ultrafast lasers addressed drilling: isolated spots, multiple pulses, comparatively simple geometry. The gap this paper targets is continuous grooving: the beam removes material while it travels, so heat accumulation and material flow evolve along the groove, and thermal effects are more involved than for an isolated hole. Most studies also cared only about geometric accuracy and neglected thermal damage. This paper puts two outputs side by side — groove width (the geometry you want) and the HAZ proxy width (the thermal cost you want to control) — which is the pair a real process-window optimization needs: optimizing groove width alone can select the most thermally damaging parameters, while minimizing thermal impact alone may fail to deliver the removal you need.

Another honest premise is dataset size: 100 experiments is small in machine-learning terms. The authors do not pretend otherwise; they make “is a simple network enough on a small dataset?” the research question itself, and answer it with cross-validation and baseline models rather than a single flattering correlation coefficient.

02 / The experiment: how the 100-point grid was built

The material was aluminium 3003 foil, 200 μm thick, cut into ten 50 mm × 50 mm samples, cleaned with isopropyl alcohol, with ten parallel grooves 5 mm long machined on each sample. Only two variables changed: average power from 1 to 10 W in 1 W steps (one power per sample) and scan speed from 10 to 100 mm/s in 10 mm/s steps (one speed per groove), giving exactly 10 × 10 = 100 data points. The laser was a Light Conversion Pharos system: 500 fs, 200 kHz, 1030 nm, with 10 W maximum average power corresponding to 50 μJ per pulse, focused through a 100 mm telecentric F-theta lens to a spot of about 100 μm, for a maximum pulse fluence of 0.6 J/cm² on the sample. Power stability was better than 0.5% and scanner speed accuracy better than 0.2%, which the authors treat as negligible here.

The measurement method sets the wording boundaries of this note. Groove width was measured with a Nikon Eclipse TS100 metallographic optical microscope and defined as the surface distance between opposite groove walls; each groove was measured at ten positions along its length, averaged and rounded to the nearest micrometre as that point’s target value, with a microscope imaging and measurement resolution of about 1 μm. Groove depth was not included: the authors explain that standard optical methods cannot clearly resolve the groove’s inner walls, so the study addresses surface groove analysis only — do not describe this “groove-width prediction” as a complete reconstruction of groove geometry.

The HAZ definition needs even more care. What the authors use is not a metallographic heat-affected zone but an “optically determined HAZ proxy width”: the total width between the outer boundaries of the visible discoloration band flanking the groove, including the groove itself; when no discoloration was detectable beyond the groove edge, the value was set equal to the groove width by convention. Discoloration is a visible indicator of thermal effects, but optical surface observation cannot fully evaluate subsurface microstructural change — the discoloration-band width, the microstructural heat-affected zone and the zone of degraded mechanical properties are three different quantities, and this model predicts only the first.

Two endpoints build intuition. At the lowest power and highest speed (1 W, 100 mm/s) the smallest groove width, 37 μm, was obtained with no discoloration detectable beyond the groove edge — narrower than the roughly 100 μm spot, which the authors explain with the Gaussian beam profile: only the beam centre exceeds aluminium’s ablation threshold (about 100 mJ/cm²) and removes material, while the remaining energy dissipates into the surroundings without producing a distinct thermal footprint. At the highest power (10 W), groove width reached 72 μm with a HAZ proxy width of 110 μm at 100 mm/s, and 112 μm with a HAZ proxy width of 140 μm at 10 mm/s: slower scanning deposits more energy per unit length, so removal and heat spreading grow together. The cover image is the 10 W, 80 mm/s case: groove width 71 μm, HAZ proxy width 120 μm, with fairly uniform groove edges but a more irregular discoloration boundary, attributed by the authors to uneven heat diffusion from material-structure variation and secondary heating by the ablation plasma.

03 / The model: why a small 2-10-5-2 network is enough

The network structure can be stated in one sentence: 2 inputs (average power, scan speed) → hidden layer 1 (10 neurons, hyperbolic tangent sigmoid transfer) → hidden layer 2 (5 neurons, the same transfer) → 2 outputs (groove width, HAZ proxy width, linear transfer), implemented in MATLAB R2021b’s Neural Network Toolbox as a feed-forward backpropagation model trained with the Levenberg–Marquardt algorithm, mean squared error as the performance function, default min–max normalization of inputs and outputs, and Nguyen–Widrow initialization of weights and biases.

The training setup is equally restrained: the 100 data points were randomly split into 70% training, 15% validation and 15% testing; the validation set monitored training and set the early-stopping point, while the test set never updated network weights. Training allowed up to 1000 epochs, a minimum performance gradient of 1.0 × 10⁻⁷, and stopping after six consecutive validation checks without improvement; in practice the minimum validation mean squared error was reached at epoch 11, training stopped after epoch 17, and the epoch-11 parameters were retained. Training, validation and test errors were comparable, with no pronounced divergence — which the authors read as evidence against severe overfitting, though a 15-point test set inherently limits how strong that conclusion can be.

Why not a bigger network? The authors’ argument is overfitting risk on a small dataset: the more complex the structure, the easier it is to memorize noise in 100 points instead of learning the parameter-to-outcome trend. The architecture was selected by trial — comparing hidden-layer counts, neuron counts and transfer functions — balancing predictive performance against structural simplicity, landing on 2-10-5-2. That “just enough” selection logic carries directly into process software: a prediction module is not necessarily better when bigger; its complexity should match the scale of the experimental data, and validation and testing must stay independent.

04 / Results and baselines: the error structure behind R = 0.95

The headline result first. The test-set overall correlation coefficient was R ≈ 0.95, split into 0.88 for groove width and 0.96 for the HAZ proxy width; the validation set gave an overall R ≈ 0.96 (0.89 and 0.93 per output) and the training set about 0.96. Five-fold cross-validation gave R = 0.94 ± 0.02 and MAPE of 8.01 ± 0.84%, consistent with the single-split results — the conclusion does not hinge on one particular random split.

The absolute error magnitudes matter more for engineering use than the correlation does. Mean absolute percentage errors were about 11% for groove width and about 15% for the HAZ proxy width; mean signed errors were about −3% and −5%, a slight tendency to underestimate. Residuals showed no clear systematic dependence on either input, which the authors read as an absence of systematic bias; but the errors come from three inseparable sources: model approximation error, fluctuation inherent in the ablation process itself, and measurement uncertainty (irregular groove edges, fuzzy discoloration boundaries). In other words, 11–15% is not a single label of “model inaccuracy” but the combined uncertainty of the whole experiment–measurement–modelling chain.

The baseline models answer “why a neural network at all”. On the same dataset, linear regression reached average correlations of 0.74 for groove width and 0.66 for the HAZ proxy width; second-order polynomial regression reached 0.88 and 0.81 — both below the neural network’s 0.95. The authors’ conclusion is that the relationship between power, speed and the two outputs is nonlinear, and neither a linear nor a polynomial function describes it fully, so the neural network’s win follows from the nonlinearity of the relationship itself rather than from parameter count. The cross-literature comparison stays measured: the authors acknowledge models with higher accuracy (one QFN cutting study reported about 1.5% test error, but with a limited input-variation range and few extreme values) and others with lower accuracy (one nanosecond-laser drilling HAZ study reported R of about 0.52 and 0.85), placing this model at a reasonable level among models trained on real measured data.

05 / Lessons for process software — and the boundaries

Three lessons carry directly into laser-processing software design. First, predict outputs in pairs: groove width together with the thermal proxy width lets you trade “is removal sufficient?” against “is thermal damage acceptable?” and find low-thermal-impact conditions; a geometry-only model can silently select the most thermally costly parameters. Second, keep inputs restrained and outputs measurable: two inputs, two optically measurable outputs and 100 experiments already yield a usable interpolation model — a realistic route for small-batch process development: build the model from a small experimental grid, then use it to screen parameters and cut trial-and-error, rather than demanding big data from the start. Third, ship the error band with the prediction: the 11–15% error range, the slight underestimation tendency and the interpolation-only applicability should be part of the parameter-recommendation interface, not footnotes hidden in a paper.

Boundary statement (consistent with the paper; please preserve it when citing). Licence and version: the paper is open access under CC BY 4.0; this note is an original interpretation with short quotations and does not reproduce the paper’s experimental data tables; the cover is Fig. 2 of the paper, resized only, with source and licence in the caption. Experimental boundaries: aluminium 3003 foil (200 μm) only, one 1030 nm / 500 fs / 200 kHz laser system only, and only inside the 1–10 W by 10–100 mm/s grid — the model is an interpolation tool, and changing material, laser system or parameter range requires fresh experimental validation and, if needed, retraining. Measurement boundaries: the HAZ width is an optical discoloration proxy width, not a microstructural heat-affected-zone width; groove depth was not measured and is not predicted; each target value averages ten positions, and point-to-point fluctuation is not modelled. Statistical boundaries: the test set holds only 15 points and five-fold cross-validation gives MAPE 8.01 ± 0.84%; a high correlation is not causation, nor a guarantee of stability under production variation. JUYAO has not replicated this on its own equipment. This is a research interpretation, not a process-recipe commitment.

PRIMARY SOURCE DOSSIER

Paper Title
Machine Learning-Based Prediction of Ablation Groove Geometry and Heat-Affected Zone Formation in Femtosecond Laser Micromachining of Aluminum
Authors
Mateusz Tański, Robert Barbucha, Marek Kocik, Todor Petrov, Mostefa Mohamed-Seghir
Journal
Materials, 2026, Vol. 19, Issue 14, 3028
Published
2026-07-14
DOI
10.3390/ma19143028
Licence
Creative Commons Attribution 4.0 International (CC BY 4.0)
Original URL
Open original paper ↗